nc://optmem

OptMem Memory: an append-only memory on your own disk, with a decay tree that keeps recent memories verbatim and compresses old ones into summaries. A Nodecode rewrite of rarf/optmem-hermes-plugin by rarf.

The model gets eight tools — optmem_note, optmem_recall, optmem_wake, optmem_nap, optmem_zoom, optmem_forget, optmem_config and optmem_import — and a /optmem command. Every request carries the store's wake digest as one optmem block: the newest memories verbatim, older ones as their one-line compressions. No key, no service, no network: keeping a memory costs no tokens.

Get it

Tell the agent "add the optmem cell", or run nodecode add nc://optmem.

Set it up

Nothing is required — the store is ~/.nodecode/optmem/, created when the cell starts. The optmem section of config.jsonc (or /settings) only changes where it lives and how it reads:

{
  "optmem": {
    "store": "~/notes/optmem",   // optional: the folder holding LOG.txt and TREE/
    "recall_mode": "auto",       // auto, regex or bm25
    "wake_lines": 96,            // how much the digest every request carries prints
    "auto_nap": false            // compress one block every ten finished turns
  }
}

A member of the wrong shape is refused at start, naming the member, and (restart-cells "nodecode-optmem") tries again after you fix it.

Use it

Remember that the deploy window is Thursday nights.

The model calls optmem_note, and the note is on disk for good — one line, one atomic fact, older notes never overwritten. Later:

What did we decide about the deploy window?

It calls optmem_recall, which searches every memory ever written: the original memo regex for pattern-like queries, ranked BM25 for prose, and auto to choose between them. optmem_zoom opens a compressed block back into its halves, down to the raw memories; optmem_forget drops a summary, never a memory.

/optmem prints the store, how many memories it holds and what compression is pending; /optmem deploy window searches it from the shell.

The decay tree

Memories are never rewritten. Each block of two, four, eight… consecutive memories can be compressed into one line of at most 280 bytes — a nap — and a block is built from its two halves, smallest first. The digest that rides your requests is a cover of that tree, so the last memories arrive verbatim while the oldest arrive as their compressions, within the wake_lines budget. A digest is complete only when every block it names has a summary; when one is missing, optmem_wake says so and hands the model the next nap's prompt.

Nothing runs in the background by default: the model compresses when the digest asks. auto_nap compresses one pending block every ten finished turns with a local extractive summary — no model call, nothing invented.

The store

LOG.txt holds one 320-byte record per memory (#<id> <date> <text>, space padded), and TREE/<size> holds the summaries of the blocks of that size at 288 bytes each. That is the format Victor Taelin's OptMem memo tool writes, so a log written here stays readable by it.

Upstream design: VictorTaelin/OptMem. This cell is the Hermes integration by rarf rewritten for Nodecode; see NOTICE. MIT, see LICENSE.